Certified Splits: Calibrated Significance for Sub-Beam Targets in Microwave Azimuth Super-Resolution

A radar or microwave radiometer cannot see two objects as separate when they sit closer together than roughly one beamwidth. The limit is about wavelength divided by antenna size, and on a car or a satellite the antenna cannot be made larger. Neural networks now "super-resolve" these sensors. They sharpen a blurred point source to half its beamwidth, or create virtual antenna elements from raw signals. When such a network shows two peaks where the raw data showed one, nothing tells the user whether the split is real or an artifact of noise. This proposal treats every sub-beam split as a statistical claim, in the way particle and cosmic-ray physicists treat a bump in a histogram. Each split gets a p-value that accounts for the search over all positions (the "look-elsewhere" effect). The p-value is computed by simulating the sensor's measured antenna pattern and noise, not by trusting the network. We then check on real hardware that these p-values mean what they say.

The opportunity is that the pieces already exist separately. Theory says how much signal-to-noise ratio is needed to resolve closely spaced sources. Hardware testbeds can place real targets at known separations. Physicists have developed trials-factor corrections for small-count anomaly searches. Nobody has combined them to answer one operational question: at what separation and signal strength can a microwave sensor's super-resolved output be trusted, and with what false-alarm rate?

Core Claim

Deciding whether a single beam contains one target or two is a model-order hypothesis test, so it can be given a calibrated error rate. The test should run on the raw measurements through a physically measured forward model. The super-resolution method, whether network or classical, only proposes candidate splits. Because the test never uses the network's output as evidence, its validity does not depend on the network. The network affects only power, meaning how small a separation gets detected. Two things become measurable for the first time on real microwave hardware. The first is the empirical false-split rate of existing super-resolvers, which no paper in this area reports. The second is the empirical minimum certifiable separation as a function of signal-to-noise ratio, which can be compared directly against the theoretical scaling law of Computational resolution limit: a theory towards super-resolution.

That paper proves, for one-dimensional band-limited measurements with Gaussian noise, that the minimum resolvable separation of $n$ sources obeys

$$\mathcal{D}_{num} \approx \frac{1}{\Omega}\left(\frac{\sigma}{m_{min}}\right)^{\frac{1}{2n-2}},$$

where $\Omega$ is the aperture's spatial-frequency cutoff, $\sigma$ the noise standard deviation, and $m_{min}$ the weakest source amplitude. For a pair of targets ($n=2$) this predicts that log separation falls with slope $-1/2$ against log signal-to-noise ratio. This is a sharp, falsifiable prediction that has never been tested with a real antenna pattern having sidelobes, discrete array sampling, and receiver thermal noise.

Research Program

The test statistic is a generalized likelihood ratio. Let $y$ be the raw measurement, either complex ADC samples for FMCW radar (frequency-modulated continuous-wave radar, the standard automotive type) or visibilities for an aperture-synthesis radiometer. Let $A_\theta$ be the forward operator that maps a set of point sources with parameters $\theta$ to noiseless measurements, built from the measured antenna pattern. For a candidate region, we compare the best $k$-source fit to the best $(k+1)$-source fit:

$$\Lambda = \min_{\theta_k} \lVert y - A_{\theta_k} \rVert^2_{\Sigma^{-1}} - \min_{\theta_{k+1}} \lVert y - A_{\theta_{k+1}} \rVert^2_{\Sigma^{-1}},$$

where $\Sigma$ is the receiver noise covariance. Two complications rule out textbook chi-square thresholds. When two sources merge, the extra source's position is unidentifiable under the null hypothesis, so Wilks' theorem fails. The search also scans every range–azimuth cell. We therefore calibrate $\Lambda$ by parametric bootstrap: we simulate the fitted $k$-source scene through the forward model many times with resampled noise and record the maximum $\Lambda$ over the whole searched field. This mirrors the $10^6$-trial Monte Carlo that Telescope Array's anisotropy analysis used to turn a $5.1\sigma$ local excess into a $3.4\sigma$ global one. That paper motivates the method: it shows how badly local significance overstates evidence when counts are small and the search is wide.

The super-resolvers act only as proposers of candidate splits. We use three: ADC-SR from Azimuth Super-Resolution for FMCW Radar in Autonomous Driving, the residual-CNN visibility extension of Visibility Extension of 1-D Aperture Synthesis by a Residual CNN, and the Poisson maximum-likelihood deconvolution of Fast Resolution Enhancement for Real Beam Mapping. Their peaks seed the $(k+1)$-source optimization, which matters because that likelihood is non-convex.

The experimental-validation direction enters at the forward model, and this is where we expect the work to be hardest. The bootstrap is only as honest as $A_\theta$. A mismatch between the assumed and true antenna pattern produces structured residuals that look like an extra source. The validation literature repeatedly finds that idealized models miss physical detail. Kim and Lee's rebar study found an oscillatory tail absent from homogeneous simulations. The bellows buckling study traced model divergence to manufacturing tolerances. We therefore measure each sensor's pattern and inter-element phase errors in an anechoic chamber, then quantify how deliberate pattern errors inflate the false-split rate. The output is an error budget: how accurately the pattern must be known for a nominal $\alpha=0.01$ to remain below 0.02 in practice.

Baselines

The naive reference point is the raw beam. A split is declared only when two peaks are separated in the un-super-resolved azimuth profile. This defines the Rayleigh-limited separation that any method must beat at equal false-split rate.

The strongest existing competitors are the super-resolvers used as their authors use them, with a split declared whenever the output shows two local maxima above a constant-false-alarm-rate threshold. ADC-SR is the strongest on radar. It reports PSNR 39.6 on Pitt-Radar for its Hybrid-SR variant and raises RADDet detection mAP from 0.764 to 0.802 (Li et al.). On radiometry the strongest is the visibility-extension ResCNN, which narrowed a physical noise source's half-power beamwidth from 1.26° to 0.58° and separated two chamber sources (Zhao et al.). For false-split rate or calibrated significance of any of these methods: no reported result found on Lacuna. We retrain each on the same simulated-plus-real training split, with the same compute and the same hyperparameter-search budget, and we tune their detection thresholds on held-out data to hit the same nominal false-split rate as our test. Beating them means a smaller certified separation at matched empirical false-split rate, or a false-split rate that stays at nominal when theirs drifts as signal-to-noise ratio falls.

The strongest classical competitor is model-order selection using the MUSIC-type singular-value-thresholding algorithm of Liu and Zhang, together with information-criterion order estimates (AIC/MDL) on the same raw data. It runs on the same measurements with no training. The key ablation replaces the bootstrap-calibrated threshold with an uncalibrated one, either a fixed $\chi^2$ quantile or a local-only bootstrap without the maximum over the field. This isolates whether look-elsewhere calibration, rather than the likelihood fit, produces any gain in validity. A second ablation swaps the proposer between MUSIC and each network while holding the test fixed. Our claim predicts that validity stays unchanged across proposers and only power differs.

Why not just…?

Why not just put a CFAR threshold on the super-resolved map? CFAR (constant false alarm rate) detection assumes the map's background statistics are those of noise passed through the sensor. A network output has learned, input-dependent statistics, and hallucinated peaks are correlated with real structure. A threshold tuned at one signal-to-noise ratio therefore need not hold at another. This is exactly what our matched-rate comparison measures. If CFAR on ADC-SR output stays calibrated across the full signal-to-noise sweep, our contribution shrinks to the forward-model error budget, and we will report that.

Why not just use deep ensembles or conformal prediction on the network output? Both quantify variability of the reconstructed image, not the probability that a second source exists. Conformal intervals on pixel values do not control the rate of declaring a nonexistent target, and they need ground-truth scenes from the deployment distribution, which Hu et al.'s multi-band work shows do not exist at satellite scale. We include a split-conformal baseline on the hardware set, where ground truth is known by construction, so the comparison is empirical rather than argued.

Why not just use classical model-order selection with MUSIC, and skip the networks? This is the real threat, and the design accounts for it. If MUSIC with a bootstrap-calibrated test matches the network-proposed test in power, the honest conclusion is that networks add nothing to certifiable resolution, only to image appearance. That would itself be a useful negative result for the PSNR-driven literature. The proposer ablation settles it directly.

Proposed Experiments

The decisive dataset is a controlled hardware collection, since public benchmarks lack known sub-beam ground truth. On radar, we use the same TI AWR1843BOOST 77 GHz, 12-virtual-element sensor used for Pitt-Radar. Pairs of trihedral corner reflectors sit on a precision rail at separations from 0.1 to 1.5 beamwidths. Attenuators and range set the signal-to-noise ratio over roughly 25 dB, and single-reflector trials at the same positions serve as nulls. On radiometry, we reproduce Zhao et al.'s one-dimensional aperture-synthesis chamber setup with paired noise sources under the same sweep. Simulated data come from HY-2B-derived visibility scenes, following Zhao et al., and from synthetic FMCW scenes. Pitt-Radar, RADDet, and the Chaotianmen bridge boat data serve as field tests, measuring how often each method declares splits where our test does not.

The first check is calibration. Across thousands of single-target trials, the empirical false-split rate of our test should fall within binomial error of nominal at every signal-to-noise level, while the threshold-tuned network baselines are expected to drift. The second check is the scaling law. On paired-target trials, we fit the separation achieving 50% detection at $\alpha=0.01$ against signal-to-noise ratio and test whether the log–log slope is $-1/2$, as the Liu–Zhang bound predicts for $n=2$. A three-reflector series probes the $n=3$ exponent of $-1/4$. Systematic deviation would quantify what sidelobes and discrete sampling add beyond the idealized theory. The third check is the pattern-error budget: injecting known amplitude and phase errors into $A_\theta$ and recording false-split inflation. The fourth is the proposer ablation described above. The project succeeds if calibration holds on hardware and the scaling is confirmed or cleanly refuted. It is equally informative if the networks turn out to buy no certifiable resolution beyond MUSIC.